Huge Opus 5.5 Leaks + Cheaper? Qwen 4, Kimi K3.1, MiniMax M3.1 & Step 5 Preview! AI News

The state of the story

The working headline remains the same: Huge Opus 5.5 Leaks + Cheaper? Qwen 4, Kimi K3.1, MiniMax M3.1 & Step 5 Preview! AI News. The current story is not a single confirmed launch. It is a cluster of signals around Anthropic’s rumored Claude Opus 5.5, StepFun’s already visible Step 5 Preview, MiniMax’s apparent M3.1 preparations, and community speculation around Qwen 4 and Kimi K3.1.

The most important distinction is confidence. Step 5 Preview has crossed from rumor into product availability through StepFun’s API, with open weights promised later. Opus 5.5, by contrast, is still being described through leaks, routing strings and second-hand market chatter, not through an Anthropic model card or official launch page. MiniMax M3.1 sits in the middle: there are repository and community signals, plus earlier management comments, but not a public production launch. Qwen 4 and Kimi K3.1 remain the thinnest part of the bundle: the market is watching them because the surrounding release cadence is accelerating, not because either has been formally unveiled.

Opus 5.5: the leak is specific, but still a leak

The Anthropic thread centers on claims that a model labeled Claude Opus 5.5, with an internal or early-access codename claude-wafer-eap, is being tested ahead of a possible Tuesday, September 22, 2026 release [2]. That timing matters because it would place Anthropic’s next flagship response just days after a weekend of community discussion about rival frontier models and Chinese lab launches [1].

The alleged pricing is the eye-catcher. Reports describe a rate of $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and cache writes at $5 per million tokens [2]. If accurate, that would be a meaningful cut from the cited Opus 5 level of $5 input and $25 output per million tokens, effectively a 20% decrease in the headline input and output rates [2]. For developers running agentic workflows, that difference is not cosmetic: long-horizon coding, retrieval, browser use and multi-step tool execution can multiply token burn quickly.

But the caution flag is large. APIMaster’s September 21 analysis says Opus 5.5 was not released at publication time and that there was no official model ID, model card, price sheet or Anthropic announcement for it [2]. OrcaRouter’s September 20 write-up reaches a similar conclusion: the claude-wafer-eap claim is precise enough to verify, but still unconfirmed and dependent on limited sourcing [3]. That makes the right editorial posture “high-signal rumor,” not “launch.”

Why a cheaper Opus would matter

If the reported price is real, Anthropic would be making a competitive move in two directions at once. First, it would defend the premium Claude line against rival frontier models on quality. Second, it would make the economics of persistent agents less punishing. The HelloBro subject brief frames Opus 5.5 as a possible replacement for the earlier Opus 5.2 track and as the most likely near-term Anthropic release among several rumored systems [1].

The practical question is not just whether $4/$20 pricing appears on a web page. It is whether the model can deliver the same or better task success with equal or fewer tokens. A nominal 20% price reduction can disappear if a model reasons longer, calls tools more often or generates verbose intermediate traces. Conversely, a modest price cut can compound if reliability improves and developers need fewer retries.

That is why the alleged model behavior matters almost as much as the price. The rumor stream emphasizes long-horizon execution, coding-heavy tasks and complex visual or 3D generation workflows [1]. Those are exactly the areas where frontier labs are trying to convert benchmark gains into paid usage: not a prettier chatbot answer, but a system that can keep a project thread alive for hours.

Step 5 Preview is the most concrete release in the bundle

Among the models in this news cluster, StepFun’s Step 5 Preview is the least speculative. DeAI reported on September 21 that StepFun had shipped the Step 5 Preview API, describing it as a 600B-parameter sparse Mixture-of-Experts model with 27B active parameters per token, a 1M-token context window, and a list price of $1 per million input tokens and $2.70 per million output tokens [4]. The same report says StepFun plans to release open weights on October 15, 2026, while noting that no license had been named [4].

That last caveat is crucial. “Open weights” can mean anything from broadly usable commercial deployment to a research-only package with practical restrictions. For builders, the October 15 date is not the end of the diligence process; it is the beginning. The license, serving recipe, quantization friendliness, throughput and actual hosted alternatives will determine whether Step 5 becomes a practical open-weight agent model or simply a strong hosted API with a future download promise.

Still, StepFun’s positioning is clear. The company is not only claiming intelligence; it is claiming better cost per task. In a market where frontier performance is expensive, a 600B-total model that activates only 27B parameters per token is designed to change the economics of serving. Even if it does not beat the absolute top model in every benchmark, it can matter if it delivers enough capability at a much lower operating cost.

MiniMax M3.1: repository smoke before launch fire

MiniMax M3.1 is another near-release candidate, but the public evidence is indirect. Juya AI Daily reported on September 20 that strings for MiniMax-M3.1 appeared across multiple files in MiniMax’s open-source Code CLI, including model-list, parameter and persistence-related areas [5]. The same report notes context and output configurations ranging across large values, plus multiple thinking-effort levels, while stressing that M3.1 had not yet entered the actual built-in model list [5].

AGI HUNT’s September 20 digest separately tracked the same theme, saying an X user had pointed to a MiniMax M3.1 string in a recent model-selection.test.ts file in the MiniMax-Code repository and that MiniMax had not commented publicly [7]. Taken together, the evidence looks like late-stage plumbing rather than a user-facing launch.

The expectation around MiniMax is shaped by its broader roadmap. The HelloBro brief notes that MiniMax leadership previously described M3.1, M3 Pro and H3.1 as close to completion, with M3.1 focused less on raw parameter spectacle and more on stability, output quality, inference efficiency and agent generalization [1]. That framing fits the market moment: teams want models that are not just smarter in demos, but cheaper and steadier in production loops.

Qwen 4 and Kimi K3.1: watchlist, not confirmation

The Qwen and Kimi parts of the headline require the most restraint. DealCan’s September 20 tech roundup lists Alibaba’s Qwen4 series and Moonshot’s Kimi K3.1 among the domestic Chinese models expected around the Mid-Autumn and National Day holiday period, alongside Step 5 Preview, MiniMax M3.1 and others [8]. That makes them part of the same release-wave narrative, but it does not make either model launched.

For Alibaba, the confirmed activity is elsewhere. Qwen’s team released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model reported to cut average lag from 2.8 seconds to 2.3 seconds across 60 languages [6]. That release does not prove Qwen 4 is imminent, but it does show Alibaba continuing to ship specialized model updates during the same window.

For Moonshot, the signal is more teaser-like. Juya AI Daily reported that Kimi’s verified Zhihu account posted digits of pi beginning after “3.1,” which the community interpreted as a possible Kimi K3.1 hint; the same report says the post did not directly mention K3.1, did not provide timing or capability details, and was later deleted [5]. That is classic pre-launch rumor fuel, not confirmation.

The bigger pattern: capability is being repackaged as economics

The real story is not simply “more models.” It is that frontier AI competition is shifting from raw benchmark bragging to price-performance, context length, active-parameter efficiency and agent reliability. Anthropic’s rumored Opus 5.5 price cut would address the premium end of the market. StepFun’s Step 5 Preview attacks the cost-per-task angle with sparse activation. MiniMax appears to be preparing a reliability-and-efficiency iteration. Alibaba and Moonshot are being watched because both have the scale and user base to turn a version-number update into market pressure.

For now, the editorial bottom line is clear: Step 5 Preview is live as an API; Opus 5.5 is a specific but unconfirmed leak; MiniMax M3.1 looks close but unreleased; Qwen 4 and Kimi K3.1 remain watchlist items. If Tuesday brings an Anthropic launch, the story becomes a pricing war. If it does not, the story is still a useful snapshot of where the next AI cycle is heading: bigger contexts, lower per-task costs, and models built for agents rather than chat alone.

Sources from the last 72 hours

1. [1]Huge Opus 5.5 Leaks + Cheaper? Qwen 4, Kimi K3.1, MiniMax M3.1 & Step 5 Preview! AI NewsSep 21, 2026, 6:15 AM UTC
2. [2]Claude Opus 5.5: Anthropic Is Skipping a Version to Fight GPT-6 | APIMaster.AISep 21, 2026, 12:00 AM UTC
3. [3]Claude Opus 5.5 泄露：一个名为“claude-wafer-eap”的检查点，以及传闻中的周二发布Sep 20, 2026, 12:00 AM UTC
4. [4]StepFun's Step 5 Preview API ships; open weights promised October 15Sep 21, 2026, 12:00 AM UTC
5. [5]橘鸦 AI 日报 — Juya AI DailySep 19, 2026, 4:00 PM UTC
6. [6]Alibaba Qwen Team Releases Qwen3.8-LiveTranslate: A Real-Time Interpretation Model That Cuts Average Lag to 2.3 Seconds Across 60 LanguagesSep 20, 2026, 12:00 AM UTC
7. [7]AGI HUNT · AI News Daily 2026-09-20 — Today's AI HighlightsSep 20, 2026, 12:00 AM UTC
8. [8]Tech Express | September 20, 2026 Over a Dozen Large Models Set to Launch Around China's Double Holiday - DealCanSep 20, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.